v1.0 — Cost-optimization lint for multi-agent Claude Code harnesses

plz-save-token helps you build well-formed multi-agent Claude Code harnesses. Catches misallocated team compositions and burning cost patterns — not the multi-agent design itself. Companion to harness-diagnostic.

What ships in v1.0

• 10 categories × 31 sub-patterns, registered only when N≥2 independent external sources confirm
• 4 modes: Pre-flight · Authoring · Audit · Hook
• 3 headline artifacts:
  • references/task_to_model_matrix.md — 24-row matrix (Python / Haiku / Sonnet / Opus with citations)
  • scripts/model_selector.py — LLM-0 decision tree CLI
  • references/optimal_team_composition.md — 8 workload patterns × 5 cost-optimization patterns
• scripts/audit.py — 9 static rules across your ~/.claude (standard library only)
• scripts/estimate_cost.py — single-source pricing module
• scripts/hook_check.py — runtime guard (PreToolUse + UserPromptSubmit, never blocks)
• Community channels: issue templates for baseline submission, pattern proposal, and quality regression reports

Self-consistency

Tool passes its own audit — R1-R9 → FAIL 0, WARN 0, PASS 4, N/A 5. meta_skill: true frontmatter exempts the meta-skill from R4/R9 (context-bloat rules).

Quick start

git clone https://github.com/epoko77-ai/plz-save-token
cd plz-save-token
python3 scripts/audit.py                                      # catalog-wide
python3 scripts/model_selector.py --task "your task description"

Standard library only. Python 3.10+. No pip install.

Caveats (honest)

• Single-operator baseline (one 27-harness catalog). Inbound diagnostics welcome.
• Claude Code-specific. LangChain/CrewAI/AutoGen need adapters (planned).
• Pricing is point-in-time (2026-05-14). Manual update when Anthropic prices change.
• Quality regression risk: downgrade recommendations need golden-task validation. See README.

Full README: https://github.com/epoko77-ai/plz-save-token